[Opencode Go] Frequent jitter / possible truncation when calling opencode-go/muse-spark-1.2 via Codex (opencodex proxy, openai-chat)

Description

Calling opencode-go/muse-spark-1.2-contributor (also muse-spark-1.2) through Codex → opencodex (openai-chat → https://opencode.ai/zen/go/v1/chat/completions) frequently jitters: Codex spins, retries 1/5…5/5, then stream disconnected before completion: upstream stream ended mid tool call without a terminal signal — possible truncation. Pure chat (echo hi) is stable; tool-calling turns (List files, cat …, any tool_calls with ~119 tools) reproduce at high rate.
Possible root cause: upstream stream for this model ends as tool_calls → usage/cost with no finish_reason and no [DONE]. src/adapters/openai-chat.ts:1689 treats !sawFinish && pendingToolCalls>0 as truncation even when pendingUsage already arrived. deepseek-v4-flash under identical conditions correctly sends finish_reason:"tool_calls" and succeeds.
[ocx:openai-chat:stream-truncated] {"finishReason":null,"hadUsage":true,"pendingToolCalls":1}

Plugins

opencodex (on codex)

OpenCode version

1.18.11

Steps to reproduce

1. Configure opencode-go (baseUrl https://opencode.ai/zen/go/v1, any modelContextWindows)
2. ocx sync (catalog 35 models, muse-spark present) and opencodex at 127.0.0.1:10100 (openai_base_url injected)
3. Via Codex (ChatGPT account, codex 0.147.0, opencodex 2.28.0):
   codex exec -m opencode-go/muse-spark-1.2-contributor "List files"
   or: codex exec -m opencode-go/muse-spark-1.2-contributor "Run: cat | head -n 3"
4. Enable ocx debug provider on to see stream-truncated on each tool turn.

Screenshot and/or share link

No response

Operating System

CentOS7

Terminal

Linux Terminal